Comprehensive Evaluation of Artificial Intelligence Models for Diagnosis of Multiple Sclerosis Using Information from Retinal Layers Multicenter OCT Images

Abstract Multiple sclerosis (MS) is a chronic inflammatory disease that affects the central nervous system. Optical coherence tomography (OCT) is a retinal imaging technology with great promise as a possible MS biomarker. Unlike other ophthalmologic diseases, the variations in shape of raw cross-sectional OCTs in MS are subtle and not differentiable from healthy controls (HCs). More detailed information like thickness of particular layers of retinal tissues or surface of individual retinal boundaries are more appropriate discriminators for this purpose. Artificial Intelligence (AI) has demonstrated a robust performance in feature extraction and classification of retinal OCTs in different ophthalmologic diseases using OCTs. We explore a comprehensive range of AI models including (1) feature extraction with autoencoder (AE) and shallow networks for classification, (2) classification with deep networks designed from scratch, and (3) fine-tuning of pretrained networks (as a generic model of the visual world) for this specific application. We also investigate different input data including thickness and surfaces of different retinal layers to find the most representative data for discrimination of MS. Moreover, channel-wise combination and mosaicing of multiple inputs are examined to find the better merging model. To address interpretability requirement of AI models in clinical applications, the visualized contribution of each input data to the classification performance is shown using occlusion sensitivity and Grad-CAM approaches. The data used in this study includes 38 HC and 78 MS eyes from two independent public and local datasets. The effectiveness and generalizability of the classification methods are demonstrated by testing the network on these independent datasets. The most discriminative topology for classification, utilizing the proposed deep network designed from scratch, is determined when the inputs consist of a channel-wise combination of the thicknesses of the three layers of the retina, namely the retinal fiber layer (RNFL), ganglion cell and inner plexiform layer (GCIP), and inner nuclear layer (INL). This structure resulted in balanced-accuracy of 97.3, specificity of 97.3, recall 97.4%, and g-mean of 97.3% in discrimination of MS and HC OCTs.